Accurately diagnosing SPNs that are malignant is vital for the effective detection and management of lung cancer. While a range of models have been designed to date with the goal of distinguishing between SPNs that are benign and malignant based on certain biomarkers [19–21], it is necessary to further stratify these analyses according to SPN size owing to the high degree of variability in terms of malignancy rates among SPNs of different sizes [22]. While there have been some specific predictive models designed for the evaluation of small SPNs [8, 19, 23], there has been little effort to date focused on incorporating radiomics data into these models.
Variables such as clinical and tumor morphological characteristics are often incorporated into clinicoradiological predictive models aimed at differentiating benign SPNs from nodules that are malignant [8, 19]. The most common CT features that are considered characteristic of SPN malignancy include a larger diameter, lobulation, spiculation, and CT bronchial sign [8, 19, 23]. Accordingly, a more traditional predictive model based on clinical and tumor CT findings was designed in the present study, yielding respective AUC values of 0.853 and 0.816 in the training and testing cohorts. These AUC values align well with those of other predictive models for small SPNs that have been reported in past studies (0.744–0.878) [8, 19]. However, these traditional CT features fail to offer any insight into the detailed internal structural properties of target tumors. Moreover, the identification of these features is often based on the experience of the radiologists who evaluate patient imaging results such that they are prone to a high risk of bias.
The radiomics method entails the processing of medical images to extract high-dimensional quantitative data. This technique can enable the characterization of tumor microscopic features related to cellular, molecular, or gene expression patterns. A growing body of evidence supports the application of radiomics to the differential diagnosis and prognostic assessment of several tumor types [12–14].
Here, a CT radiomics-based model was developed that was capable of distinguishing between benign and malignant small SPNs. This model was based on a combination of the radiomics scores and clinical model established herein, and it exhibited an AUC value superior to that of the clinical model in the training (0.957 vs. 0.853, P = 0.021) and testing (0.943 vs. 0.816, P < 0.001) cohorts. These data offer clear evidence for the ability of the radiomics score to significantly improve diagnostic performance relative to that associated with traditional clinical and radiological findings. The resultant nomogram can generate a direct predictive score for each small SPN, with this score corresponding to a predicted probability that can aid in clinical decision-making efforts.
Predictive models developed to evaluate small SPNs in prior reports determined that CEA levels were significantly related to the risk of malignancy [8, 23]. One meta-analysis further demonstrated that CEA exhibited good diagnostic performance when used to distinguish between benign and malignant PNs [24]. In the present study, however, no relationship was detected between tumor marker levels and the malignancy status of small SPNs. These discrepant results may be attributable to sample size limitations.
There are some limitations to the present study. For one, as a retrospective study, there is a high risk of selective bias. Secondly, this was a single-center study such that prospective multi-center validation will be vital. Thirdly, some patient data at baseline was not balanced between the training and testing cohorts, potentially contributing to a greater risk of bias. However, both cohorts exhibited similarly high AUC values exceeding 0.9, suggesting a high degree of stability for this predictive model. Finally, as a radiomics approach, the reproducibility of this analytical strategy and its potential for standardization are limited, constraining the potential clinical application of this model.